Estimating Off‐Fault Deformation Using Convolutional Neural Networks Trained on Experimental Strike‐Slip Faults

Scaled physical experiments using crustal analog materials can simulate off‐fault deformation near strike‐slip faults and provide direct observations of deformation partitioning during strike‐slip fault evolution. Using experimental fault maps and off‐fault deformation labels, we trained and tested four 2D Convolutional Neural Networks (CNNs) that can estimate off‐fault deformation from active strike‐slip fault trace maps. Experiments with different analog materials, loading rates, and basal boundary conditions provide a range of conditions that impact fault evolution and the partitioning of strain as off‐fault deformation at each stage of fault development. Each CNN trained on fault maps from either wet kaolin, poured sand, or sedimented sand achieve an accuracy >90% when predicting the off‐fault deformation of an unseen data set. The CNN trained on the combination of all three data sets achieves lower accuracy of 86% because of differences of early strike‐slip fault patterns in the sedimented sand from the wet kaolin and poured sand. CNN models can be applied to crustal fault maps because they are trained on physical experiments that are scaled to represent upper crustal deformation. Of the four CNNs, the CNN trained on wet kaolin and poured sand performs best on crustal fault maps that have independent geologic estimates of off‐fault deformation. With this validation, the experiment‐trained CNN could provide off‐fault deformation estimates of crustal strike‐slip faults, as long as loading conditions are similar to those applied in the experiments. Off‐fault deformation estimates for crustal faults for which we do not yet have geologic estimates can improve seismic hazard assessment.